小花自我迭代 (HuaNiu Enhanced)

小花专用自我迭代技能 - 基于 self-improving-agent 增强，集成 OpenClaw 工作流、MEMORY.md、百度千帆、看想做找四部曲。专为国内部署优化。

Security Scan

The skill's files, scripts, and runtime instructions are coherent with its stated purpose (self‑improvement / learning capture for OpenClaw) and do not request unrelated credentials or perform suspicious network/download actions.

Purpose & Capability

Name/description match the included files and instructions: the SKILL.md, hooks, and scripts implement reminders, error detection, and a helper to extract learnings into new skills. References to China-friendly tooling (Baidu, Ollama) are documented as integration options rather than required credentials. There is a minor metadata mismatch (package _meta.json says 'self-improving-agent' v2.0.0 while registry entry is 'xiaohua-self-improving' v1.0.0) but this looks like a fork/branding difference rather than malicious behavior.

Instruction Scope

Runtime instructions and hooks operate on workspace files (MEMORY.md, .learnings/, WORKFLOW_AUTO.md) and inject short reminders during bootstrap or after tool use. The error-detector reads the CLAUDE_TOOL_OUTPUT environment variable (expected for a PostToolUse hook). The docs mention OpenClaw session APIs (sessions_history, sessions_send, sessions_spawn) as available capabilities — these are examples and not used by the provided hook code, but they are powerful (can read/send cross-session data) and worth noting before enabling cross-session features.

Install Mechanism

No downloader or package install is declared (instruction-only with bundled scripts and hooks). All included scripts are plain shell/JS/TS files; there are no remote fetches, URL shorteners, or archive extraction behaviors in the manifest. This is a low-risk install model assuming you review and opt-in to enable hooks.

Credentials

The skill declares no required environment variables or credentials. Scripts use CLAUDE_TOOL_OUTPUT (a runtime tool-output variable) which is expected for an error-detection hook. There are no requests for AWS/GitHub/other unrelated keys or secrets in the files.

Persistence & Privilege

always:false and user-invocable:true (default). Hooks/scripts are opt‑in: enabling them requires explicit user config (openclaw hooks enable or adding entries to settings files). The scripts run with the same permissions as the user/agent process — standard for hooks — and they do not modify other skills' configurations.

Assessment

This skill appears internally consistent and implements short, opt‑in reminders and error-detection hooks for an OpenClaw workspace. Before enabling it: 1) review the hook scripts (hooks/openclaw/handler.{js,ts}, scripts/*.sh) to confirm you accept their behavior; they will run with the same user permissions as the agent. 2) If you enable hooks globally (~/.claude or ~/.openclaw settings), prefer a sandbox workspace first to verify outputs. 3) Note extract-skill.sh writes to a relative ./skills directory but validates paths to avoid absolute or '..' writes — still confirm the target directory. 4) The skill references cross-session APIs in docs; those are powerful — only enable inter-session features if you trust the environment. 5) There are no credentials requested by this skill; services like Baidu/OLLAMA are mentioned as optional integrations and would require separate configuration. If you want higher assurance, run the scripts in a controlled environment (non-production user) before enabling them in your main workspace.

🌸 小花定制版: 基于 self-improving-agent fork，专为 OpenClaw + 国内环境优化。集成 MEMORY.md、看想做找工作流、百度千帆 API、生产环境实战模式。

快速开始

# 安装
npx clawhub install xiaohua-self-improving

# 使用后记录
# 1. 错误 → .learnings/ERRORS.md
# 2. 修正 → .learnings/LEARNINGS.md
# 3. 功能请求 → .learnings/FEATURE_REQUESTS.md
# 4. 通用知识 → 提升到 MEMORY.md

核心增强

1. OpenClaw 原生集成

• ✅ 自动加载到 ~/.openclaw/workspace/
• ✅ 与 AGENTS.md、SOUL.md、TOOLS.md 联动
• ✅ 支持 memory/YYYY-MM-DD.md 日常记录
• ✅ 与 WORKFLOW_AUTO.md 看想做找流程整合

2. MEMORY.md 提升工作流

.learnings/LEARNINGS.md  (详细上下文)
         ↓
   MEMORY.md  (团队通用知识)
         ↓
   SOUL.md  (行为准则/人格)
         ↓
   TOOLS.md  (工具使用模式)

3. 国内友好配置

• 🔍 搜索: 百度千帆 API (baidu-search 技能)
• 🧠 记忆: local-file-rag-basic 或 scripts/memory_search.py
• 🤖 模型: 阿里云千帆 / Ollama 本地
• 🌐 部署: 无需 VPN，全国内服务

4. 生产环境模式

• 上下文窗口修复流程 (配置→验证→重启)
• 技能安装测试规范 (装一个→测一个→记一个)
• 错误追踪模板 (标准化 ERR-XXX 格式)
• 日常记忆追加协议 (APPEND only)

文件结构

~/.openclaw/workspace/
├── AGENTS.md              # 多智能体工作流
├── SOUL.md                # 人格/行为准则
├── TOOLS.md               # 工具能力/坑点
├── MEMORY.md              # 长期记忆 (主会话)
├── memory/                # 日常记忆 (YYYY-MM-DD.md)
├── WORKFLOW_AUTO.md       # 看想做找四部曲
├── HEARTBEAT.md           # 状态监控
├── USER.md                # 主人偏好
└── .learnings/            # 本技能日志
    ├── LEARNINGS.md
    ├── ERRORS.md
    └── FEATURE_REQUESTS.md

使用示例

记录错误

### [ERR-20260302-002] ChromaDB 不兼容 Python 3.14

**Logged**: 2026-03-02T15:45:00+08:00
**Priority**: high
**Status**: workaround
**Area**: dependencies

### Summary
ChromaDB 1.5.2 与 Python 3.14 不兼容 (Pydantic v1 问题)

### Error

pydantic.v1.errors.ConfigError: unable to infer type for attribute "chroma_server_nofile"

### Context
- Python: 3.14.0
- ChromaDB: 1.5.2
- Pydantic v1 已停止维护

### Suggested Fix
1. 临时：使用 local-file-rag-basic
2. 长期：等 ChromaDB 更新兼容版本
3. 替代：FAISS 或 LanceDB

### Metadata
- Reproducible: yes
- Tags: chromadb, python3.14, pydantic
- See Also: LRN-20260302-003

提升到 MEMORY.md

## 主人的教导记录
### 2026-03-02
- Tool Call ID 格式：9 位字符 (A-Z, a-z, 0-9)，禁用特殊符号
- 上下文窗口：以实际 API 返回为准 (阿里云千帆编码版 32K，非文档 131K)
- PowerShell JSON：复杂管道易错，直接写文件更安全

看想做找工作流

## 看 (Look)
- 读取 WORKFLOW_AUTO.md、MEMORY.md、HEARTBEAT.md
- 扫描主人消息语气和情绪

## 想 (Think)
- 分析意图和优先级
- 匹配响应风格

## 找 (Find)
- 搜索 MEMORY.md 相关知识
- web_search / baidu-search 查找信息
- 检查 .learnings/ 避免重复错误

## 做 (Do)
- 执行任务
- 记录结果到 .learnings/
- 提升到 MEMORY.md (如适用)
- 验证输出质量

最佳实践

1. 立即记录 - 上下文最新鲜
2. 具体详细 - 未来智能体需要快速理解
3. 包含复现步骤 - 特别是错误
4. 链接相关文件 - 便于修复
5. 建议具体方案 - 不只是"调查"
6. 使用一致分类 - 便于筛选
7. 积极提升 - 有疑问就加到 MEMORY.md
8. 定期回顾 - 过时的学习会贬值
9. 工作流自动化 - 遵循看想做找
10. 国内部署 - 用百度 API、本地模型、国内工具

实战模式

模式 1: 日常记忆追加

更新 memory/YYYY-MM-DD.md 时:

• 只追加 (APPEND)，绝不覆盖
• 用时间戳标题：## ✅ 下午技术修复完成 (2026-03-02 16:00)
• 包含文件清单便于追溯

模式 2: 错误追踪

使用 .learnings/ERRORS.md 模板:

### [ERR-YYYYMMDD-XXX] skill_name

**Logged**: ISO 时间戳
**Priority**: high/medium/low
**Status**: pending/workaround/fixed
**Area**: dependencies/config/skill/other

### Summary
一句话描述

### Error

实际错误信息

### Context
- 环境细节
- 尝试了什么

### Suggested Fix
临时方案或永久修复

### Metadata
- Reproducible: yes/no
- Tags: 逗号分隔
- See Also: LRN-XXX (如相关)

模式 3: 上下文窗口修复

当模型上下文与文档不符时:

1. 检查实际 API 返回值
2. 更新 openclaw.json 匹配现实
3. 重启网关：Stop-Process → 等 2 秒 → openclaw gateway
4. 记录到 MEMORY.md 作为技术修复

模式 4: 技能安装

• 一个一个装: clawhub install 只接受单个 slug
• 装完测试: 验证功能再依赖
• 记录 bug: 坏了就加到 .learnings/ERRORS.md

与其他技能协作

技能	协作方式
baidu-search	国内搜索替代 Brave/Tavily
local-file-rag-basic	本地记忆检索 (ChromaDB 不兼容时)
clawhub	技能安装/发布/更新
ollama-local	本地模型部署 (无需 API)
deep-research-pro	深度研究 (需配置 Tavily API)

已知限制

• ❌ baidu-academic 技能不存在 (clawhub 无此技能)
• ❌ ChromaDB 不兼容 Python 3.14+ (Pydantic v1 问题)
• ❌ local-file-rag-basic 可能有 bug (CodeChunker is not defined)
• ⚠️ Tavily 搜索国内信息效果差 (返回英文结果)

更新日志

v1.0.0 (2026-03-02)

• ✅ 基于 self-improving-agent fork
• ✅ OpenClaw 原生集成
• ✅ MEMORY.md 提升工作流
• ✅ 百度千帆 API 配置
• ✅ 看想做找自动化
• ✅ 生产环境实战模式
• ✅ 国内部署优化
• ✅ 错误追踪模板
• ✅ 上下文窗口修复流程

参考

• 原始技能：https://github.com/peterskoett/self-improving-agent
• OpenClaw 文档：https://docs.openclaw.ai
• ClawHub: https://clawhub.com
• 小花文档：C:\Users\qwe12\.openclaw\workspace\

---

🌸 小花出品，必属精品 - 持续迭代，越用越聪明